Open Stack

> -----Original Message-----
> From: Sahid Orentino Ferdjaoui [mailto:sferdjao at redhat.com]
> Sent: Tuesday, September 26, 2017 1:46 PM
> To: OpenStack Development Mailing List (not for usage questions)
> <openstack-dev at lists.openstack.org>
> Subject: Re: [openstack-dev] vGPUs support for Nova
> 
> On Mon, Sep 25, 2017 at 04:59:04PM +0000, Jianghua Wang wrote:
> > Sahid,
> >
> > Just share some background. XenServer doesn't expose vGPUs as mdev or
> > pci devices.
> 
> That does not make any sense. There is physical device (PCI) which
> provides functions (vGPUs). These functions are exposed through mdev
> framework. What you need is the mdev UUID related to a specific vGPU
> and I'm sure that XenServer is going to expose it. Something which
> XenServer may not expose is the NUMA node where the physical device is
> plugged on but in such situation you could still use sysfs.
[Mooney, Sean K] this is implementation specific. Amd support virtualizing
There gpu using sriov http://www.amd.com/Documents/Multiuser-GPU-White-Paper.pdf
In that case you can use the existing pci pass-through support without any modification.
For intel and nvidia gpus we need speficic hypervisor support as the device partitioning
Is done in the host gpu driver rather than via sirov. There are two level of abstraction
That we must keep separate. 1 how does the hardware support configuration and enumeration
Of the virutalised resources (amd in hardware via sriov, intel/nvidia via driver/software manager). 
2 how does the hypervisor report the vgpus to openstack and other clients.

In the amd case I would not expect any hypervisor to have mdevs associated with 
The sriov vf as that is not the virtualization model they have implemented.
In the intel gvt case yes you will have mdevs but the virtual gpus are not 
Represented on the pci bus so we should not model them as pci deveices.

Some more comments below.
> 
> > I proposed a spec about one year ago to make fake pci devices so that
> > we can use the existing PCI mechanism to cover vGPUs. But that's not
> a
> > good design and got strongly objection. After that, we switched to
> use
> > the resource providers by following the advice from the core team.
> >
> > Regards,
> > Jianghua
> >
> > -----Original Message-----
> > From: Sahid Orentino Ferdjaoui [mailto:sferdjao at redhat.com]
> > Sent: Monday, September 25, 2017 11:01 PM
> > To: OpenStack Development Mailing List (not for usage questions)
> > <openstack-dev at lists.openstack.org>
> > Subject: Re: [openstack-dev] vGPUs support for Nova
> >
> > On Mon, Sep 25, 2017 at 09:29:25AM -0500, Matt Riedemann wrote:
> > > On 9/25/2017 5:40 AM, Jay Pipes wrote:
> > > > On 09/25/2017 05:39 AM, Sahid Orentino Ferdjaoui wrote:
> > > > > There is a desire to expose the vGPUs resources on top of
> > > > > Resource Provider which is probably the path we should be going
> > > > > in the long term. I was not there for the last PTG and you
> > > > > probably already made a decision about moving in that direction
> > > > > anyway. My personal feeling is that it is premature.
> > > > >
> > > > > The nested Resource Provider work is not yet feature-complete
> > > > > and requires more reviewer attention. If we continue in the
> > > > > direction of Resource Provider, it will need at least 2 more
> > > > > releases to expose the vGPUs feature and that without the
> > > > > support of NUMA, and with the feeling of pushing something
> which is not stable/production-ready.
[Mooney, Sean K] Not all gpus have numam affinity. Intel integrated gpus do not. they have
Dedicated edram on the processor die so there memory accesses never leave
The processor package sot they do not have numa affinity. I would assume the
Same is true for amd integrated gpus so only descreet gpus will have numa affinity.
> > > > >
> > > > > It's seems safer to first have the Resource Provider work well
> > > > > finalized/stabilized to be production-ready. Then on top of
> > > > > something stable we could start to migrate our current virt
> > > > > specific features like NUMA, CPU Pinning, Huge Pages and
> finally PCI devices.
> > > > >
> > > > > I'm talking about PCI devices in general because I think we
> > > > > should implement the vGPU on top of our /pci framework which is
> > > > > production ready and provides the support of NUMA.
> > > > >
> > > > > The hardware vendors building their drivers using mdev and the
This is vendor specific intel uses mdevs for intel GVT(kvmgt/xengt).
Amd uses sriov and does not use mdevs it uses sriov http://www.amd.com/Documents/Multiuser-GPU-White-Paper.pdf.

Amd is simple because you just do a pci passthough of the sriov-vf and your are done.
No explcit support need in the hypervisior.

Looking at https://images.nvidia.com/content/grid/pdf/GRID-vGPU-User-Guide.pdf section
5.3.3.2. when we query a specific physical gpu for the support paramaters via xen it reports the pci address of that
Physical gpu and part of the responce

[root at xenserver ~]# xe pgpu-param-list uuid=f2607117-5b4c-d6cc-3900-00bf712e33f4
uuid ( RO) : f2607117-5b4c-d6cc-3900-00bf712e33f4
 vendor-name ( RO): NVIDIA Corporation
 device-name ( RO): GK104GL [GRID K2]
 gpu-group-uuid ( RW): f4662c69-412c-abc5-6d02-f74b7703cccd
 gpu-group-name-label ( RO): GRID K2 Socket 0
 host-uuid ( RO): d9eb9118-a5c5-49fb-970e-80e6a8f7ff98
 host-name-label ( RO): xenserver-vgx-test (VM IPs 10.31.223.0-49,
 dom0 .96, OOB .97)
 pci-id ( RO): 0000:08:00.0
 dependencies (SRO):
 other-config (MRW):
 supported-VGPU-types ( RO): a724b756-d108-4c9f-0ea3-8f3a1553bfbc;
 63d9d912-3454-b020-8519-58dedb3b0117; 0bdf4715-e035-19c3-a57d-5ead20b3e5cd;
 a7838abe-0d73-1918-7d29-fd361d3e411f
 enabled-VGPU-types (SRW): a724b756-d108-4c9f-0ea3-8f3a1553bfbc;
 63d9d912-3454-b020-8519-58dedb3b0117; 0bdf4715-e035-19c3-a57d-5ead20b3e5cd;
 a7838abe-0d73-1918-7d29-fd361d3

looking at section 5.2 I would appear that for nvida we create vgpus using the group rather
then the parent gpu as follows
xe vgpu-create vm-uuid=e71afda4-53f4-3a1b-6c92-a364a7f619c2 
gpu-group-uuid=be825ba2-01d7-8d51-9780-f82cfaa64924 
vgpu-typeuuid=3f318889-7508-c9fd-7134-003d4d05ae56b73cbd30-096f-8a9a-523e-a800062f4ca7

that makes sense because they handel allocation to a specific gpu based on a policy 
(pack vs spread) as described in sect 5.3. the groups are auto created by xen based
On vendor-id and product id of the nvidia gpu but based on the group name I would guess
It also groups by socket which they may be conflating with numa node. In either case
We could use the pci address and sysfs to confirm the numa affintity if needed.

In the intel case https://github.com/01org/gvt-linux/wiki/GVTg_Setup_Guide#41-linux-guest-setup
Looking at section 5.3 in the kvm case we create memdevs by echoing a uuid to a sysfs file that is a child of
The intergraged gpu on the pci bus e.g. 
echo "a297db4a-f4c2-11e6-90f6-d3b88d6c9525" > "/sys/bus/pci/devices/0000:00:02.0/ mdev_supported_types/i915-GVTg_V4_4/create" 
and then do a pci passthough of that memdev

e.g. section 5.6.1
/bin/bash -x
    /usr/bin/qemu-system-x86_64 \
    -m 2048 -smp 2 -M pc \
    -name gvt-g-guest \
    -hda /home/img/ubuntu-1.qcow2  \
    -bios /usr/bin/bios.bin -enable-kvm \
    -net nic,macaddr=00:A1:00:00:00:1A -net tap,script=/etc/qemu-ifup \
    -vga qxl \
    -k en-us \
    -serial stdio \
    -vnc :1 \
    -machine kernel_irqchip=on \
    -global PIIX4_PM.disable_s3=1 -global PIIX4_PM.disable_s4=1 \
    -cpu host -usb -usbdevice tablet \
    -device vfio-pci,sysfsdev=/sys/bus/pci/devices/0000:00:02.0/a297db4a-f4c2-11e6-90f6-d3b88d6c9525,rombar=0

So in the intel case the vgpu is directly a child of a specific pci device but in the nvidia case it’s a child of a pool of pci devices.

> > > > > /pci framework currently understand only SRIOV but on a quick
> > > > > glance it does not seem complicated to make it support mdev.
> > > > >
> > > > > In the /pci framework we will have to:
> > > > >
> > > > > * Update the PciDevice object fields to accept NULL value for
> > > > >    'address' and add new field 'uuid'
I think the PFs should still have pciadresses
The vfs on the other hand will be created dynmaicaly so
If they were tracked in this table perhaps the uuid would be more appropriate.
In the amd case though the vgpus really do have pci addresses so you cant assume uuids 
> > > > > * Update PciRequest to handle a new tag like 'vgpu_types'
If we where to do this I would prefer if this was added into the capabilities
Dictionary we are introducing in the nic feature based scheduling spec.
We intentionally made that generic so that in addition to holding network capabilities
It could have gpu or fpga or other-fancy-acllerator if needed though traits will ultimately 
Be the end goal. See https://review.openstack.org/#/c/451777/26/nova/pci/stats.py

> > > > > * Update PciDeviceStats to also maintain pool of vGPUs
> > > > >
> > > > > The operators will have to create alias(-es) and configure
> > > > > flavors. Basically most of the logic is already implemented and
> > > > > the method 'consume_request' is going to select the right vGPUs
> > > > > according the request.
> > > > >
> > > > > In /virt we will have to:
> > > > >
> > > > > * Update the field 'pci_passthrough_devices' to also include
> > > > > GPUs
> > > > >    devices.
> > > > > * Update attach/detach PCI device to handle vGPUs
> > > > >
> > > > > We have a few people interested in working on it, so we could
> > > > > certainly make this feature available for Queen.
> > > > >
> > > > > I can take the lead updating/implementing the PCI and libvirt
> > > > > driver part, I'm sure Jianghua Wang will be happy to take the
> > > > > lead for the virt XenServer part.
> > > > >
> > > > > And I trust Jay, Stephen and Sylvain to follow the
> developments.
> > > >
> > > > I understand the desire to get something in to Nova to support
> > > > vGPUs, and I understand that the existing /pci modules represent
> > > > the fastest/cheapest way to get there.
> > > >
> > > > I won't block you from making any of the above changes, Sahid.
> > > > I'll even do my best to review them. However, I will be primarily
> > > > focusing this cycle on getting the nested resource providers work
> > > > feature-complete for (at least) SR-IOV PF/VF devices.
> > > >
> > > > The decision of whether to allow an approach that adds more to
> the
> > > > existing /pci module is ultimately Matt's.
> > > >
> > > > Best,
> > > > -jay
> > > >
> > > >
> __________________________________________________________________
> > > > __ ______ OpenStack Development Mailing List (not for usage
> > > > questions)
> > > > Unsubscribe:
> > > > OpenStack-dev-request at lists.openstack.org?subject:unsubscribe
> > > > http://lists.openstack.org/cgi-bin/mailman/listinfo/openstack-dev
> > >
> > > Nested resource providers is not merged or production ready because
> > > we haven't made it a priority. We've certainly talked about it and
> > > Jay has had patches proposed for several releases now though.
> > >
> > > Building vGPU support into the existing framework, which only a
> > > couple of people understand - certainly not me, might be a
> > > short-term gain but is just more technical debt we have to pay off
> > > later, and delays any focus on nested resource providers for the
> wider team.
> > >
> > > At the Queens PTG it was abundantly clear that many features are
> > > dependent on nested resource providers, including several
> > > networking-related features like bandwidth-based scheduling.
> > >
> > > The priorities for placement/scheduler in Queens are:
> > >
> > > 1. Dan Smith's migration allocations cleanup.
> > > 2. Alternative hosts for reschedules with cells v2.
> > > 3. Nested resource providers.
> > >
> > > All of these are in progress and need review.
> > >
> > > I personally don't think we should abandon the plan to implement
> > > vGPU support with nested resource providers without first seeing
> any
> > > code changes for it as a proof of concept. It also sounds like we
> > > have a pretty simple staggered plan for rolling out vGPU support so
> > > it's not very detailed to start. The virt driver reports vGPU
> > > inventory and we decorate the details later with traits (which Alex
> Xu is working on and needs review).
> > >
> > > Sahid, you could certainly implement a separate proof of concept
> and
> > > make that available if the nested resource providers-based change
> > > hits major issues or goes far too long and has too much risk, then
> > > we have a contingency plan at least. But I don't expect that to get
> > > review priority and you'd have to accept that it might not get
> > > merged since we want to use nested resource providers.
> >
> > That seems to be fair, I understand your desire to make the
> implementation on Resource Provider a priority and I'm with you. In
> general my preference is to do not stop progress on virt features
> because we have a new "product" on-going.
> >
> > > Either way we are going to need solid functional testing and that
> > > functional testing should be written against the API as much as
> > > possible so that it works regardless of the backend implementation
> > > of the feature. One of the big things we failed at in Pike was not
> > > doing enough functional testing of move operations with claims in
> > > the scheduler earlier in the cycle. That all came in late and we're
> still fixing bugs as a result.
> >
> > It's very true and most of the time we are asking our users to be
> beta-testers, that is one more reason why my preference is for a real
> deprecation phase.
> >
> > > If we can get started early on the functional testing for vGPUs,
> > > then work both implementations in parallel, we should be able to
> > > retain the functional tests and determine which implementation we
> > > ultimately need to go with probably sometime in the second
> milestone.
> > >
> > > --
> > >
> > > Thanks,
> > >
> > > Matt
> > >
> > >
> ____________________________________________________________________
> > > __ ____ OpenStack Development Mailing List (not for usage
> questions)
> > > Unsubscribe:
> > > OpenStack-dev-request at lists.openstack.org?subject:unsubscribe
> > > http://lists.openstack.org/cgi-bin/mailman/listinfo/openstack-dev
> >
> >
> ______________________________________________________________________
> > ____ OpenStack Development Mailing List (not for usage questions)
> > Unsubscribe:
> > OpenStack-dev-request at lists.openstack.org?subject:unsubscribe
> > http://lists.openstack.org/cgi-bin/mailman/listinfo/openstack-dev
> >
> ______________________________________________________________________
> > ____ OpenStack Development Mailing List (not for usage questions)
> > Unsubscribe:
> > OpenStack-dev-request at lists.openstack.org?subject:unsubscribe
> > http://lists.openstack.org/cgi-bin/mailman/listinfo/openstack-dev
> 
> _______________________________________________________________________
> ___
> OpenStack Development Mailing List (not for usage questions)
> Unsubscribe: OpenStack-dev-
> request at lists.openstack.org?subject:unsubscribe
> http://lists.openstack.org/cgi-bin/mailman/listinfo/openstack-dev